llama : add DRY sampler (#9702)

* sampling : add DRY sampler (post-refactor) * DRY: Trying to fix coauthors, removed unneeded line * DRY: Fixed redundant code * DRY: Fixed crash issue due to DRY being in chain but uninitialized --------- Co-authored-by: l3utterfly <gc.pthzfoldr@gmail.com> Co-authored-by: pi6am <34464159+pi6am@users.noreply.github.com>
2024-10-25 10:07:34 -06:00 · 2024-10-25 10:07:34 -06:00 · ff252ea48e
commit ff252ea48e
parent d80fb71f8b
17 changed files with 713 additions and 63 deletions
--- a/examples/main/README.md
+++ b/examples/main/README.md
@ -187,6 +187,30 @@ Use the `--no-penalize-nl` option to disable newline penalization when applying

 Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`

+### DRY Repetition Penalty
+
+DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repetition in generated text even across long contexts by penalizing tokens based on their recent usage patterns (original [PR link](https://github.com/oobabooga/text-generation-webui/pull/5677)).
+
+- `--dry-multiplier N`: Set the DRY sampling multiplier (default: 0.0, 0.0 = disabled).
+- `--dry-base N`: Set the DRY sampling base value (default: 1.75).
+- `--dry-allowed-length N`: Set the allowed length for DRY sampling (default: 2).
+- `--dry-penalty-last-n N`: Set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size).
+- `--dry-sequence-breaker STRING`: Add a sequence breaker for DRY sampling. Can be used more than once to add multiple sequence breakers. Using this clears out the default breakers, which consist of: `['\n', ':', '"', '*']`. If the string `"none"` is supplied, no sequence breakers are used.
+
+The `dry-multiplier` option controls the strength of the DRY sampling effect. A value of 0.0 disables DRY sampling, while higher values increase its influence. A typical recommended value is 0.8.
+
+The `dry-base` option sets the base value for the exponential penalty calculation in DRY sampling. Higher values lead to more aggressive penalization of repetitions.
+
+The `dry-allowed-length` option sets the maximum length of repeated sequences that will not be penalized. Repetitions shorter than or equal to this length are not penalized, allowing for natural repetitions of short phrases or common words.
+
+The `dry-penalty-last-n` option controls how many recent tokens to consider when applying the DRY penalty. A value of -1 considers the entire context. Use a positive value to limit the consideration to a specific number of recent tokens.
+
+The `dry-sequence-breaker` option adds a single sequence breaker and can be used more than once to specify multiple sequence breakers. Sequence breakers interrupt sequence matching and break the input into parts where matching can be applied.
+
+DRY sampling provides more nuanced control over text generation, particularly for reducing long-range repetitions and maintaining global coherence.
+
+Example usage: `--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --dry-penalty-last-n -1 --dry-sequence-breaker "—" --dry-sequence-breaker "##"`
+
 ### Top-K Sampling

 -   `--top-k N`: Limit the next token selection to the K most probable tokens (default: 40).
--- a/examples/server/README.md
+++ b/examples/server/README.md
@ -114,6 +114,11 @@ The project is under active development, and we are [looking for feedback and co
 | `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
 | `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
 | `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
+| `--dry-multiplier N` | DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--dry-base N` | DRY sampling base value (default: 1.75) |
+| `--dry-allowed-length N` | allowed length for DRY sampling (default: 2) |
+| `--dry-penalty-last-n N` | DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
+| `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers (`['\n', ':', '"', '*']`) in the process; use `"none"` to not use any sequence breakers
 | `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
 | `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
 | `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
@ -369,6 +374,16 @@ node index.js

    `frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled.

+    `dry_multiplier`: Set the DRY (Don't Repeat Yourself) repetition penalty multiplier. Default: `0.0`, which is disabled.
+
+    `dry_base`: Set the DRY repetition penalty base value. Default: `1.75`
+
+    `dry_allowed_length`: Tokens that extend repetition beyond this receive exponentially increasing penalty: multiplier * base ^ (length of repeating sequence before token - allowed length). Default: `2`
+
+    `dry_penalty_last_n`: How many tokens to scan for repetitions. Default: `-1`, where `0` is disabled and `-1` is context size.
+
+    `dry_sequence_breakers`: Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\n', ':', '"', '*']`
+
    `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.

    `mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`
--- a/examples/server/public/index-new.html
+++ b/examples/server/public/index-new.html
@ -40,6 +40,10 @@
      repeat_last_n: 0, // 0 = disable penalty, -1 = context size
      repeat_penalty: 1.0, // 1.0 = disabled
      penalize_nl: false, // true only useful for infinite completion
+      dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
+      dry_base: 1.75,     // 0.0 = disabled
+      dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well
+      dry_penalty_last_n: -1, // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
      top_k: 0, // <= 0 to use vocab size
      top_p: 1.0, // 1.0 = disabled
      min_p: 0.05, // 0 = disabled; recommended for non-english: ~ 0.4
@ -833,13 +837,17 @@ return html`
        <fieldset class="params">
          ${IntField({ label: "Top-K", title: "Limits the selection of the next token to the K most probable tokens. 1 means no randomness = greedy sampling. If set to 0, it means the entire vocabulary size is considered.", max: 100, min: 0, step: 1, name: "top_k", value: params.value.top_k })}
          ${IntField({ label: "Penalize Last N", title: "The last n tokens that are taken into account to penalise repetitions. A value of 0 means that this function is deactivated and -1 means that the entire size of the context is taken into account.", max: 2048, min: 0, step: 16, name: "repeat_last_n", value: params.value.repeat_last_n })}
-          ${FloatField({ label: "Top-P", title: "Limits the selection of the next token to a subset of tokens whose combined probability reaches a threshold value P = top-P. If set to 1, it means the entire vocabulary size is considered.", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
          ${FloatField({ label: "Presence Penalty", title: "A penalty that is applied if certain tokens appear repeatedly in the generated text. A higher value leads to fewer repetitions.", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
-          ${FloatField({ label: "TFS-Z", title: "Activates tail-free sampling, a method used to limit the prediction of tokens that are too frequent. The parameter z controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
          ${FloatField({ label: "Frequency Penalty", title: "A penalty that is applied based on the frequency with which certain tokens occur in the training data set. A higher value results in rare tokens being favoured.", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
+          ${FloatField({ label: "Top-P", title: "Limits the selection of the next token to a subset of tokens whose combined probability reaches a threshold value P = top-P. If set to 1, it means the entire vocabulary size is considered.", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
          ${FloatField({ label: "Typical-P", title: "Activates local typical sampling, a method used to limit the prediction of tokens that are atypical in the current context. The parameter p controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
          ${FloatField({ label: "XTC probability", title: "Sets the chance for token removal (checked once on sampler start)", max: 1.0, min: 0.0, name: "xtc_probability", step: 0.01, value: params.value.xtc_probability })}
          ${FloatField({ label: "XTC threshold", title: "Sets a minimum probability threshold for tokens to be removed", max: 0.5, min: 0.0, name: "xtc_threshold", step: 0.01, value: params.value.xtc_threshold })}
+          ${FloatField({ label: "DRY Penalty Multiplier", title: "Set the DRY repetition penalty multiplier. Default is 0.0, which disables DRY.", max: 5.0, min: 0.0, name: "dry_multiplier", step: 0.01, value: params.value.dry_multiplier })}
+          ${FloatField({ label: "DRY Base", title: "Set the DRY repetition penalty base value. Default is 1.75", max: 3.0, min: 1.0, name: "dry_base", step: 0.01, value: params.value.dry_base })}
+          ${IntField({ label: "DRY Allowed Length", title: "Tokens that extend repetition beyond this receive exponentially increasing penalty. Default is 2", max: 10, min: 1, step: 1, name: "dry_allowed_length", value: params.value.dry_allowed_length })}
+          ${IntField({ label: "DRY Penalty Last N", title: "How many tokens to scan for repetitions. Default is -1, where 0 is disabled and -1 is context size", max: 2048, min: -1, step: 16, name: "dry_penalty_last_n", value: params.value.dry_penalty_last_n })}
+          ${FloatField({ label: "TFS-Z", title: "Activates tail-free sampling, a method used to limit the prediction of tokens that are too frequent. The parameter z controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
          ${IntField({ label: "Min Keep", title: "If greater than 0, samplers are forced to return N possible tokens at minimum. Default is 0", max: 10, min: 0, name: "min_keep", value: params.value.min_keep })}
        </fieldset>

@ -1144,6 +1152,8 @@ document.addEventListener('DOMContentLoaded', (event) => {
    repeat_penalty: { snapValue: 1.0, snapRangeMultiplier: 4 },
    presence_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 },
    frequency_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 },
+    dry_multiplier: { snapValue: 0.0, snapRangeMultiplier: 4 },
+    dry_base: { snapValue: 1.75, snapRangeMultiplier: 4 },
  };
  // add an event listener for each slider
  Object.keys(snapSettings).forEach(sliderName => {
--- a/examples/server/public/index.html
+++ b/examples/server/public/index.html
@ -304,6 +304,10 @@
      repeat_last_n: 256, // 0 = disable penalty, -1 = context size
      repeat_penalty: 1.18, // 1.0 = disabled
      penalize_nl: false,
+      dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
+      dry_base: 1.75,     // 0.0 = disabled
+      dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well
+      dry_penalty_last_n: -1, // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
      top_k: 40, // <= 0 to use vocab size
      top_p: 0.95, // 1.0 = disabled
      min_p: 0.05, // 0 = disabled
@ -1015,6 +1019,10 @@
              ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
              ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
              ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
+              ${FloatField({ label: "DRY Penalty Multiplier", max: 5.0, min: 0.0, name: "dry_multiplier", step: 0.01, value: params.value.dry_multiplier })}
+              ${FloatField({ label: "DRY Base", max: 3.0, min: 1.0, name: "dry_base", step: 0.01, value: params.value.dry_base })}
+              ${IntField({ label: "DRY Allowed Length", max: 10, min: 2, step: 1, name: "dry_allowed_length", value: params.value.dry_allowed_length })}
+              ${IntField({ label: "DRY Penalty Last N", max: 2048, min: -1, step: 16, name: "dry_penalty_last_n", value: params.value.dry_penalty_last_n })}
              ${FloatField({ label: "XTC probability", max: 1.0, min: 0.0, name: "xtc_probability", step: 0.01, value: params.value.xtc_probability })}
              ${FloatField({ label: "XTC threshold", max: 0.5, min: 0.0, name: "xtc_threshold", step: 0.01, value: params.value.xtc_threshold })}
            </fieldset>
--- a/examples/server/public/style.css
+++ b/examples/server/public/style.css
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@ -800,35 +800,58 @@ struct server_context {
            slot.oaicompat_model = "";
        }

-        slot.params.stream             = json_value(data, "stream",            false);
-        slot.params.cache_prompt       = json_value(data, "cache_prompt",      false);
-        slot.params.n_predict          = json_value(data, "n_predict",         json_value(data, "max_tokens", default_params.n_predict));
-        slot.params.n_indent           = json_value(data, "n_indent",          default_params.n_indent);
-        slot.sparams.top_k             = json_value(data, "top_k",             default_sparams.top_k);
-        slot.sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
-        slot.sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
-        slot.sparams.xtc_probability   = json_value(data, "xtc_probability",   default_sparams.xtc_probability);
-        slot.sparams.xtc_threshold     = json_value(data, "xtc_threshold",     default_sparams.xtc_threshold);
-        slot.sparams.tfs_z             = json_value(data, "tfs_z",             default_sparams.tfs_z);
-        slot.sparams.typ_p             = json_value(data, "typical_p",         default_sparams.typ_p);
-        slot.sparams.temp              = json_value(data, "temperature",       default_sparams.temp);
-        slot.sparams.dynatemp_range    = json_value(data, "dynatemp_range",    default_sparams.dynatemp_range);
-        slot.sparams.dynatemp_exponent = json_value(data, "dynatemp_exponent", default_sparams.dynatemp_exponent);
-        slot.sparams.penalty_last_n    = json_value(data, "repeat_last_n",     default_sparams.penalty_last_n);
-        slot.sparams.penalty_repeat    = json_value(data, "repeat_penalty",    default_sparams.penalty_repeat);
-        slot.sparams.penalty_freq      = json_value(data, "frequency_penalty", default_sparams.penalty_freq);
-        slot.sparams.penalty_present   = json_value(data, "presence_penalty",  default_sparams.penalty_present);
-        slot.sparams.mirostat          = json_value(data, "mirostat",          default_sparams.mirostat);
-        slot.sparams.mirostat_tau      = json_value(data, "mirostat_tau",      default_sparams.mirostat_tau);
-        slot.sparams.mirostat_eta      = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
-        slot.sparams.penalize_nl       = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
-        slot.params.n_keep             = json_value(data, "n_keep",            default_params.n_keep);
-        slot.params.n_discard          = json_value(data, "n_discard",         default_params.n_discard);
-        slot.sparams.seed              = json_value(data, "seed",              default_sparams.seed);
-        slot.sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
-        slot.sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
-      //slot.params.t_max_prompt_ms    = json_value(data, "t_max_prompt_ms",   default_params.t_max_prompt_ms); // TODO: implement
-        slot.params.t_max_predict_ms   = json_value(data, "t_max_predict_ms",  default_params.t_max_predict_ms);
+        slot.params.stream              = json_value(data, "stream",             false);
+        slot.params.cache_prompt        = json_value(data, "cache_prompt",       false);
+        slot.params.n_predict           = json_value(data, "n_predict",          json_value(data, "max_tokens", default_params.n_predict));
+        slot.params.n_indent            = json_value(data, "n_indent",           default_params.n_indent);
+        slot.sparams.top_k              = json_value(data, "top_k",              default_sparams.top_k);
+        slot.sparams.top_p              = json_value(data, "top_p",              default_sparams.top_p);
+        slot.sparams.min_p              = json_value(data, "min_p",              default_sparams.min_p);
+        slot.sparams.xtc_probability    = json_value(data, "xtc_probability",    default_sparams.xtc_probability);
+        slot.sparams.xtc_threshold      = json_value(data, "xtc_threshold",      default_sparams.xtc_threshold);
+        slot.sparams.tfs_z              = json_value(data, "tfs_z",              default_sparams.tfs_z);
+        slot.sparams.typ_p              = json_value(data, "typical_p",          default_sparams.typ_p);
+        slot.sparams.temp               = json_value(data, "temperature",        default_sparams.temp);
+        slot.sparams.dynatemp_range     = json_value(data, "dynatemp_range",     default_sparams.dynatemp_range);
+        slot.sparams.dynatemp_exponent  = json_value(data, "dynatemp_exponent",  default_sparams.dynatemp_exponent);
+        slot.sparams.penalty_last_n     = json_value(data, "repeat_last_n",      default_sparams.penalty_last_n);
+        slot.sparams.penalty_repeat     = json_value(data, "repeat_penalty",     default_sparams.penalty_repeat);
+        slot.sparams.penalty_freq       = json_value(data, "frequency_penalty",  default_sparams.penalty_freq);
+        slot.sparams.penalty_present    = json_value(data, "presence_penalty",   default_sparams.penalty_present);
+        slot.sparams.dry_multiplier     = json_value(data, "dry_multiplier",     default_sparams.dry_multiplier);
+        slot.sparams.dry_base           = json_value(data, "dry_base",           default_sparams.dry_base);
+        slot.sparams.dry_allowed_length = json_value(data, "dry_allowed_length", default_sparams.dry_allowed_length);
+        slot.sparams.dry_penalty_last_n = json_value(data, "dry_penalty_last_n", default_sparams.dry_penalty_last_n);
+        slot.sparams.mirostat           = json_value(data, "mirostat",           default_sparams.mirostat);
+        slot.sparams.mirostat_tau       = json_value(data, "mirostat_tau",       default_sparams.mirostat_tau);
+        slot.sparams.mirostat_eta       = json_value(data, "mirostat_eta",       default_sparams.mirostat_eta);
+        slot.sparams.penalize_nl        = json_value(data, "penalize_nl",        default_sparams.penalize_nl);
+        slot.params.n_keep              = json_value(data, "n_keep",             default_params.n_keep);
+        slot.params.n_discard           = json_value(data, "n_discard",          default_params.n_discard);
+        slot.sparams.seed               = json_value(data, "seed",               default_sparams.seed);
+        slot.sparams.n_probs            = json_value(data, "n_probs",            default_sparams.n_probs);
+        slot.sparams.min_keep           = json_value(data, "min_keep",           default_sparams.min_keep);
+      //slot.params.t_max_prompt_ms     = json_value(data, "t_max_prompt_ms",    default_params.t_max_prompt_ms); // TODO: implement
+        slot.params.t_max_predict_ms    = json_value(data, "t_max_predict_ms",   default_params.t_max_predict_ms);
+
+        if (slot.sparams.dry_base < 1.0f)
+        {
+           slot.sparams.dry_base = default_sparams.dry_base;
+        }
+
+        // sequence breakers for DRY
+        {
+            // Currently, this is not compatible with TextGen WebUI, Koboldcpp and SillyTavern format
+            // Ref: https://github.com/oobabooga/text-generation-webui/blob/d1af7a41ade7bd3c3a463bfa640725edb818ebaf/extensions/openai/typing.py#L39
+
+            if (data.contains("dry_sequence_breakers")) {
+                slot.sparams.dry_sequence_breakers = json_value(data, "dry_sequence_breakers", std::vector<std::string>());
+                if (slot.sparams.dry_sequence_breakers.empty()) {
+                    send_error(task, "Error: dry_sequence_breakers must be a non-empty array of strings", ERROR_TYPE_INVALID_REQUEST);
+                    return false;
+                }
+            }
+        }

        // process "json_schema" and "grammar"
        if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) {
@ -1132,6 +1155,11 @@ struct server_context {
            {"repeat_penalty",            slot.sparams.penalty_repeat},
            {"presence_penalty",          slot.sparams.penalty_present},
            {"frequency_penalty",         slot.sparams.penalty_freq},
+            {"dry_multiplier",            slot.sparams.dry_multiplier},
+            {"dry_base",                  slot.sparams.dry_base},
+            {"dry_allowed_length",        slot.sparams.dry_allowed_length},
+            {"dry_penalty_last_n",        slot.sparams.dry_penalty_last_n},
+            {"dry_sequence_breakers",     slot.sparams.dry_sequence_breakers},
            {"mirostat",                  slot.sparams.mirostat},
            {"mirostat_tau",              slot.sparams.mirostat_tau},
            {"mirostat_eta",              slot.sparams.mirostat_eta},
@ -1970,7 +1998,6 @@ struct server_context {

                                            for (size_t i = 0; i < n_match; i++) {
                                                slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i];
-
                                                slot.n_past++;
                                            }